A collection of Digital Signal Processing notebooks with a wireless communications theme.

On the Effects of Quantisation on Model Uncertainty in Bayesian Neural Networks

Wii/GC texture encoder with palette support (png, jpeg, gif, bmp, tiff, tpl, tex0, plt0, bti)

My Tensorflow Notebook. In this notebooks I have implemented various kind of model optimisation techniques.

LoRA fine-tune and serve NVFP4 models on one DGX Spark (GB10, 128 GB UMA): text backbones via generic-family onboarding, and VLM vision towers (Pixtral end-to-end; Llama-4 training-path validated). Fused Triton dequant; runtime-LoRA and merge serving.

Applying SDLC for JPEG compression project

This Repository Stores an empirical study conducted on Meta-Llama-3-8B-Instruct Model , Running on a Nvidia T4 16Gb . Quantizing it to 4-bit NF4 to reclaim 54% VRAM and accelerate Llama-3-8B generation up to 8.2x on Edge Hardware . with Negligible Loss in Model Quality . Benchmarks Showcase the Results of the Research .

JPEG Encoding with Discrete Cosine Transform, Quantisation & Run-Length Compression

Image compression codec based on JPEG implemented in python

Transcribe meeting audio and summarises.

A research based project to optimize the graph coloring problem using the quantisation part of image compression and come up with a solution which is better than conventional algorithms for graph coloring.

Optimising train, inference and throughput of expensive ML models

Experiments in quantisation consisting of quantisation from scratch, bitsandbytes, and llama.cpp. [Assignment 4 of Advanced Natural Language Processing, IIIT-H Monsoon '24]

Simple dithering library in Rust, based on image-rs

Source code: C++, SFML, MATLAB. The aim of the project was to implement various quantization algorithms in C ++ and to compare them in terms of the quality of the obtained results and the time needed to perform the calculations. On the way I've created an app with very basic GUI.

Pytorch implementation of Self Attention Enhanced Post Training Quantisation for Diffusion Models

A complete pipeline for deep learning model optimization via teacher fine-tuning, knowledge distillation, and 4-bit quantization. Built with PyTorch, Transformers, and BitsAndBytes.

Distributed GPT-2 fine-tuning with PyTorch FSDP and BF16 mixed precision, INT8 post-training quantisation, a custom Triton quantisation kernel achieving 1.4x throughput over unfused PyTorch, and a full FP32 vs BF16 vs INT8 benchmark suite. 13 tests passing.

A friendly CLI tool for converting and uploading transformers for CTranslate2.